Table of Contents
Author’s note: Qualcomm provided travel, lodging and other accommodations associated with Dragonwing IoT Day.
Across robotics, industrial automation and 6G, Qualcomm is confronting a common systems engineering challenge in turning adaptive, artificial intelligence into reliable operational outcomes
SAN DIEGO — Consider an industrial robotic arm programmed to insert a bolt into a hole on an automotive assembly line. The robot can repeat the operation with extraordinary speed and precision, but if the equipment becomes slightly miscalibrated, it may repeatedly miss its target. A human worker would recognize the misalignment, adjust and complete the task. The robot, operating according to predefined instructions, lacks that adaptability.
Qualcomm VP and GM of Automated Driving and Robotics Anshuman Saxena used this example during a roundtable discussion at Dragonwing IoT Day to illustrate the opportunity for generalized physical AI. Rather than programming a robot to execute a specific motion under specific conditions, AI could enable it to perceive changing circumstances, reason through a problem and adjust its behavior to accomplish the intended outcome.
But this introduces a fundamental tension. Industrial systems are expected to deliver predictable, repeatable results. AI models, particularly those responsible for reasoning and planning, operate probabilistically. They provide flexibility but cannot independently guarantee correct behavior.
Reconciling those characteristics is emerging as a common systems-engineering challenge for Qualcomm not only in robotics but across industrial automation, autonomous vehicles and AI-native 6G networks. The technical implementations differ, but the underlying question is how can increasingly autonomous systems respond to unpredictable conditions without compromising the operational requirements of their environments?
Generalization versus the realities of industrial automation
For Saxena, generalized physical AI offers a potential way beyond the limitations of conventional automation. Robots can already perform repetitive tasks in highly structured environments. Generalization would allow those systems to manage greater variability without requiring engineers to explicitly program every possible scenario.
That evolution is particularly important for brownfield industrial environments where equipment is already deployed and the cost of replacing existing automation can be prohibitive. However, he also distinguished generalized intelligence from the specialized implementations industrial applications will continue to require when responding to the observation that generalized robotics tend to be “bad at everything.”
“Dragonwing is a big umbrella,” Saxena explained. “Not everything is generalized,” he continued, referencing the variety of SoC platforms encompassed by Dragonwing, as well as industrial PLCs, microcontrollers and other parts of the whole. He stressed that it is a platform capable of providing customizable solutions to a variety of industries. “Everything is there…That’s what Dragonwing is.”
A robot performing a narrowly defined task — Saxena gave the example of a robotic system frying chicken — may not require a large general-purpose model. A smaller, application-specific model, informed by more generalized reasoning or training, would be sufficient. Different systems will require different combinations of perception, planning, control and execution.
Tosh Tambe, who joined Qualcomm in May as vice president of industrial automation and manufacturing after nearly eight years heading Siemens’s SaaS strategy, approached the same challenge from a more immediate operational perspective. He emphasized human-robot collaboration rather than the prospect of generalized autonomous machines replacing factory workers. Human judgment and experience can bridge the gap between the probabilistic behavior of AI systems and the predictable outcomes required in manufacturing.
This does not mean human oversight replaces validated control systems or other engineering safeguards. Rather, people can resolve exceptions, apply contextual knowledge and perform tasks where complete automation remains technically difficult or economically unjustified. That human could, for instance, ensure a miscalibrated robotic arm still puts a bolt in a hole correctly.
His perspective also recognizes that the optimal degree of autonomy varies across industrial processes. Some applications will accommodate substantial automation; others, including certain forms of high-value aerospace manufacturing, may retain significant human involvement. The near-term opportunity is not necessarily full autonomy, but efficiently distributing work among humans, machines and software.
From individual intelligence to coordinated systems
Tambe also described the broader problem of making independently intelligent industrial assets operate as a coherent system. He estimated that autonomous mobile robots may spend only 40% to 45% of their time doing productive work. Individually capable machines can still operate inefficiently because they lack awareness of one another and their shared environment.
Qualcomm sees opportunities to address this through technologies developed for other markets. Automotive V2X communications could help robots anticipate movement around obstructed corners. Ultra-wideband positioning could improve navigation and docking. Shared contextual intelligence could help coordinate assets from different manufacturers. “These are all technologies we have,” Tambe said. “Applying them to the industrial context suddenly changes the whole solution.”
The ambition extends beyond improving individual robots. Tambe discussed work toward a common intelligence or control plane at the factory edge that would allow otherwise disconnected industrial systems to share operational context. This creates an additional challenge in that systems must coordinate adaptive decisions while preserving the timing, safety and reliability requirements of industrial operations.
It also suggests how Qualcomm’s traditional strengths in connectivity, heterogeneous computing and sensing could become more valuable when integrated at the platform level — as it’s doing with Dragonwing — rather than sold as discrete components.
The same tension in AI-native 6G
This problem has a direct parallel in telecommunications. As explored in an earlier research note, The core tension of telco AI — it’s time for a leap of faith, the transition toward autonomous networks requires building deterministic operating envelopes around probabilistic intelligence. Networks must accommodate dynamic traffic patterns, shifting radio conditions and increasingly unpredictable application requirements while preserving service availability, reliability, policy enforcement and operational accountability.
Qualcomm is applying this logic both to the use of AI in network operations and to the design of 6G itself. In its Agentic RAN Management Service, specialized AI agents are intended to analyze network conditions, generate action plans, validate proposed changes through digital twins and execute them using established network management capabilities.
At the radio-system level, Qualcomm is also designing for adaptability within prescribed operational boundaries. In an October 21 technical discussion, VP of Technology Gavin Horn described how 6G devices could dynamically adjust selected protocol behaviors in response to real-time application and radio conditions while remaining subject to network-defined policies.
The comparison with robotics is instructive. A robot must adjust to physical variability without violating safe operating constraints. An intelligent network must respond to unpredictable demand without surrendering control over resource allocation and service behavior. These are not identical forms of determinism. Wireless service assurance is often statistical, while industrial motion control can require hard real-time safeguards. But both demand architectures that constrain how AI translates into action.
Qualcomm has explicitly linked its robotics and 6G strategies, describing a layered approach that separates deterministic, low-latency control from reactive intelligence and longer-horizon reasoning.
Diversification as a common systems engineering problem
Qualcomm’s diversification strategy has historically involved extending technologies developed for smartphones into adjacent markets, including automotive, industrial IoT, PCs and now data centers. More recently, as discussed in an earlier analysis of its agentic AI strategy, Qualcomm has positioned those businesses as participants in a distributed computing architecture spanning devices, networks, the edge and cloud.
The emerging commonality may be deeper than shared silicon IP or the ability to distribute inference across heterogeneous processors. Each business increasingly requires the integration of adaptive AI with systems that have specific operational constraints. Power efficiency, sensing, connectivity, real-time execution, software orchestration and safety become interdependent engineering problems.
The challenge for Qualcomm is to translate expertise developed in one market into reusable architectural capabilities for others without assuming that the markets themselves are interchangeable. Automotive safety systems, industrial control environments and public mobile networks impose different requirements. Neither a common processor nor a generalized AI model eliminates those differences.
The underlying engineering discipline is deciding where intelligence operates, what decisions it can make, how those decisions are validated and which mechanisms constrain execution. And this engineering discipline certainly could scale across business units.
AI does not have to become deterministic to create value. But it does have to operate within an architecture capable of delivering the required outcomes, including knowing when to defer, escalate or stop. Saxena’s miscalibrated robotic arm and Tambe’s human-supervised factory offer different perspectives on how, and on what timeline, that transition might unfold. Qualcomm’s 6G work demonstrates that the same challenge extends well beyond industrial robotics.
As Qualcomm embeds AI deeper into its products, the big opportunity is disciplined systems engineering that makes autonomy both operationally reliable and economically viable. Solving for that could be a more consequential source of differentiation than the individual technologies being diversified.
For more from Qualcomm’s Dragonwing IoT Day, read my piece, “Different industries, one approach — “Qualcomm’s Dragonwing strategy for scaling industrial IoT.”
